Nature-Inspired Computing Flashcards

Question

What are two ways to update the population in EAs?

Answer 1

* Replace all with new * Merge old + new, keep best.

Answer 2

Premature convergence (bad).

Answer 3

Representation matters a lot.

Answer 4

Optimisation & machine learning.

Answer 5

* Lecture timetables * Pipe networks for water * Antenna design * Satellite orbits * Fighter aircraft strategies * Neural networks * Cancer radiotherapy treatment planning.

Answer 6

Finding best solution in a set S.

Answer 7

* Small (easy) * Huge (10³⁰ possible solutions) * Infinite (real numbers).

Answer 8

Try every possible solution (Enumeration).

Answer 9

When S is small.

Answer 10

Tractable, in P, solvable in polynomial time.

Answer 11

Intractable, not known in P, solvable in exponential time.

Answer 12

Connect all nodes in a graph with cheapest tree (no cycles).

Answer 13

Prim’s algorithm (polynomial time).

Answer 14

It becomes hard.

Answer 15

NY Tunnels water distribution optimisation.

Answer 16

1.9 × 10²⁵ possible solutions.

Answer 17

Exact algorithms guarantee best solution, approximate algorithms provide 'good enough' solutions quickly.

Answer 18

Deliver good solutions in reasonable time.

Answer 19

Time vs quality.

Answer 20

* Selection * Variation * Population update.

Answer 21

Nearly always hard.

Answer 22

[small/easy cases].

Answer 23

[practical].

Answer 24

1. Define fitness function (objective measure). 2. Generate candidate solutions. 3. Apply: * Selection (choose parents). * Mutation (small changes). * Recombination/crossover (combine parents). * Replacement (decide which solutions survive). Repeat until solutions improve.

Answer 25

* Generational: replace whole population each cycle. * Steady-state: replace only a few at a time. * Elitist: best always survives.

Answer 26

* Single-point * Multi-point * Uniform

Answer 27

* Rank-based * Roulette Wheel * Tournament

Answer 28

* Replace weakest * Replace first weaker solution

Answer 29

* Single gene * Multiple gene

Answer 30

* Population size * Number of generations * Mutation rate * Crossover rate * Tournament size

Answer 31

Must visit all cities once, return to start. Goal = shortest possible tour.

Answer 32

1. Start with random solution c. 2. Mutate → new solution m. 3. If m is as good or better → replace c with m. 4. Repeat until stopping condition.

Answer 33

Imagine fitness as height on a landscape. HC climbs 'uphill' step by step, keeping improvements.

Answer 34

As a permutation of cities (e.g., ABDEC).

Answer 35

Landscape = all solutions (S) plotted against fitness f(s).

Answer 36

All solutions reachable by one mutation.

Answer 37

* Unimodal * Plateau * Multimodal (many peaks) * Deceptive (look good but lead to poor results)

Answer 38

It gets stuck in local optima.

Answer 39

* Allow downhill moves (Local Search) * Use a population (multiple solutions explored at once)

Answer 40

Keep track of current solution c and best so far b. Explore neighbourhood → choose candidate m. Update c and b depending on policy.

Answer 41

Better than hill climbing.

Answer 42

Still often stuck in local optima.

Answer 43

Instead of one solution, keep a population.

Answer 44

1. Need selection to choose which solution(s) to mutate. 2. Can recombine multiple solutions.

Answer 45

* Avoids local optima more effectively. * Explores multiple areas of search space in parallel.

Answer 46

EAs with populations: * Mutations = local moves. * Crossovers = combine solutions → explore new areas.

Answer 47

Start with small population of random solutions. Mutate selected parents → new solutions. Evaluate new solution: * If better than worst → replace. * If worse → discard.

Answer 48

* Many variants of EAs (different selection, mutation, replacement). * Hill climbing & local search = simple but limited (get stuck). * Population-based search = much stronger, avoids local optima.

Answer 49

A method that applies selection and genetic operators to create a whole new population. ## Footnote An elitist version copies the top n best solutions unchanged to the next generation.

Answer 50

Applies operators a few times, replacing the weakest solutions immediately, leading to gradual population evolution.

Answer 51

A strategy where a new solution replaces the overall worst solution.

Answer 52

Almost random progress with no significant improvements.

Answer 53

Premature convergence, where the population gets stuck.

Answer 54

Roulette Wheel selection.

Answer 55

* Minimisation/negative values tricky * Very fit individuals can dominate.

Answer 56

A method where the population is sorted and probabilities are assigned by rank.

Answer 57

Randomly picking t individuals and choosing the best among them.

Answer 58

To determine where to focus the search in the solution space.

Answer 59

* Exploitation: refine good solutions * Exploration: search new areas.

Answer 60

* Binary strings * Integer vectors * Real values * Permutations * Trees.

Answer 61

* It shapes the fitness landscape * It determines which operators work best.

Answer 62

Changing one value randomly.

Answer 63

To swap two values, useful for permutations.

Answer 64

Split two parents at one point and swap parts.

Answer 65

A general version of crossover that can involve multiple points.

Answer 66

* Initialize random population * Select parent (tournament) * Mutate with probability * Replace worst if better * Repeat until termination.

Answer 67

* Initialize random population * Select parents (rank-based) * Pair up for crossover & mutation * Keep best old solution (elitism) * Replace rest with children. * Repeat for set number of generations.

Answer 68

It shapes the search landscape, necessitating careful selection of encoding and operators.

Answer 69

Collective smart behaviour from simple individuals ## Footnote Key ideas include no central control and emergent behaviour through local interactions.

Answer 70

* Ant Colony Optimisation (ACO) * Particle Swarm Optimisation (PSO) ## Footnote These algorithms mimic natural systems to solve complex problems.

Answer 71

* Control nest temperature * Build bridges * Find shortest path to food ## Footnote Ants exhibit complex problem-solving abilities through simple individual actions.

Answer 72

Ants choose between two bridges; over time, they prefer the shorter one due to pheromone trails ## Footnote This experiment demonstrates how pheromone accumulation leads to emergent shortest-path finding.

Answer 73

Indirect communication through the environment ## Footnote It involves actions that trigger others or leave signals influencing future behaviour.

Answer 74

* Sematonic stigmergy * Sign-based stigmergy ## Footnote Ants primarily use sign-based stigmergy by leaving pheromone trails.

Answer 75

cities/locations ## Footnote Edges represent paths between these nodes.

Answer 76

* Leaves pheromone on edges used * Evaporation reduces pheromone on all edges ## Footnote This helps in reinforcing good paths and prevents early domination.

Answer 77

* Ants * Pheromone * Evaporation * Positive feedback * Exploration vs Exploitation ## Footnote These components work together to find optimal solutions.

Answer 78

* Mimics natural self-organisation * Uses distributed problem solving * Relies on pheromone feedback ## Footnote It is particularly effective for combinatorial optimisation problems.

Answer 79

* Route planning * Scheduling * Network design ## Footnote ACO is well-suited for these types of problems due to its iterative improvement approach.

Answer 80

* Visits each city exactly once * Minimises total distance ## Footnote TSP is NP-complete and serves as a classic testbed for optimisation algorithms.

Answer 81

* Randomly place num_ant ants on num_city cities * Each ant chooses next city using probabilistic state transition rule * Update pheromone trails after all ants complete tours * Repeat until stopping condition ## Footnote ACO involves ants constructing solutions step-by-step while communicating via pheromones.

Answer 82

* Visited or not? * Local heuristic value ηᵢⱼ = 1/dᵢⱼ * Pheromone value τᵢⱼ(t) ## Footnote The transition from city i to city j depends on these factors.

Answer 83

* Each ant deposits pheromone on edges used * Amount deposited is proportional to solution quality * Pheromone evaporates ## Footnote Evaporation prevents early convergence and encourages exploration.

Answer 84

* Found optimal solutions for small problems (~30 cities) * Good results on larger problems, but not best-known * Combining ACO with local search produced world-class results ## Footnote TSP is highly static, and specialist algorithms can outperform basic ACO.

Answer 85

* Basic Ant System (AS) * Max–Min Ant System (MMAS) * Elitist Rank Ant System ## Footnote Variants differ in how pheromone updates are done.

Answer 86

FALSE ## Footnote ACO can solve any problem representable as a path in a graph.

Answer 87

* Jobs A–E with durations & deadlines * Only one machine → jobs scheduled one by one * Fitness options: average lateness, or max lateness ## Footnote Each job is a node, and ants build a sequence as a valid schedule.

Answer 88

* Minimise weight difference across bins ## Footnote Ants choose bins for each item, building the packing incrementally like a path.

Answer 89

* Transition Rule * Update Rule * Construction Graph ## Footnote ACO is competitive with evolutionary algorithms, especially on discrete optimisation.

Answer 90

* ACO = ants building solutions + pheromone learning * Transition rule = how ants choose next step * Update rule = how good solutions reinforce trails * Works great on routing, scheduling, bin packing, sequencing * Variants help avoid stagnation & improve performance * Combine with local optimisers for best results ## Footnote This summary is designed to be ADHD-friendly.

Answer 91

* Emergent intelligent behaviour * Arises from simple interactions between individuals ## Footnote Example: Slime-mold behaves like a single amoeba when food is plentiful, and merges into a multicellular “slug” when food is scarce.

Answer 92

* Flocking * Shoaling * Swarming * Formation travelling ## Footnote Flocking is the coordinated movement of animals in groups, such as birds and fish.

Answer 93

* Rapid movement of the entire group * Collective reactions to predators * Avoiding collisions * Flocks may merge or split * Tolerant of individuals joining/leaving * No leader — coordination is decentralised * Found across many species and environments ## Footnote Flocks can range from tiny groups to massive shoals up to 17 miles long.

Answer 94

* Attraction (aggregation) * Repulsion (segregation) * Neighbour mimetism * Environmental guidance ## Footnote These forces help in take-off, landing, turning, and route-keeping.

Answer 95

* Accuracy & endurance * Energy savings * Anti-predator benefits ## Footnote Example: Geese flying in a V-formation can increase range by 70%.

Answer 96

Craig Reynolds, 1987 ## Footnote Boids are simple digital “birds” that simulate flocking behaviour.

Answer 97

* Separation * Alignment * Cohesion ## Footnote These rules generate complex flocking behaviour.

Answer 98

* Spontaneously polarise * Synchronise turning * Merge if two flocks meet * Show flash expansion * Form mini-flocks if far apart ## Footnote These behaviours are demonstrated in simulations.

Answer 99

An agent perceives its environment and acts on it through sensors and actuators ## Footnote Key concepts include percept and percept sequence.

Answer 100

* Amazon Echo * Robot assistant * Self-driving car * Cybersecurity agent ## Footnote Agents can be software or physical systems.

Answer 101

Where the agent works ## Footnote Characteristics include observability, determinism, dynamism, and task types.

Answer 102

* Solve complex problems collaboratively * Learn independently * Can be cooperative or competitive * Scalable ## Footnote Used in autonomous cars, robot factories, and automated trading.

Answer 103

* Complexity * Communication overhead * Fault detection ## Footnote These challenges arise from the need for agents to work together effectively.

Answer 104

* Flat * Hierarchical * Teams * Coalitions ## Footnote These structures illustrate how agents coordinate and communicate.

Answer 105

* Centralised fault detection * Monitoring communication patterns * Identifying faulty agents ## Footnote Most research assumes homogeneous agents, which is often not the case in real systems.

Answer 106

An optimisation algorithm inspired by flocking behaviour ## Footnote Invented by Kennedy & Eberhart in 1995, covered in detail next week.

Answer 107

* Collective movement with no leader * Driven by attraction vs repulsion * Benefits include energy savings and predator avoidance * Uses Reynolds' three rules: separation, alignment, cohesion ## Footnote Sets foundation for understanding multi-agent systems and PSO.

Answer 108

An optimisation algorithm inspired by bird flocking and swarm behaviour ## Footnote PSO was introduced by Kennedy & Eberhart in 1995.

Answer 109

A candidate solution with position and velocity (but no mass) ## Footnote Particles move through the search space to find optimal solutions.

Answer 110

* Its own best position found so far (pbest) * The swarm’s best position (gbest) ## Footnote This collective knowledge helps in quicker discovery of optimal solutions.

Answer 111

* x = (-1.3220, 0.1302, 2.3125, -0.6435, -0.6435) ## Footnote Continuous problems are common in engineering optimisation tasks.

Answer 112

* Antenna design * Aircraft wings * Amplifiers * Controllers * Circuits ## Footnote PSO is widely used in various engineering optimisation tasks.

Answer 113

* Create & initialise N particles with random positions and velocities * For each timestep: * For each particle: * Update position using velocity * Update velocity ## Footnote The algorithm iteratively improves the positions of particles.

Answer 114

* Inertia * Cognitive term (personal best) * Social term (global best) ## Footnote These components guide the movement of particles in the search space.

Answer 115

Acceleration coefficients controlling the importance of personal vs global experience ## Footnote Their values affect the convergence speed and exploration of the search space.

Answer 116

* Max iterations * “Good enough” solution found * No improvement for N iterations * Swarm has converged (normalised radius ≈ 0) ## Footnote These conditions determine when the algorithm should stop running.

Answer 117

An adaptation of PSO to work with binary strings ## Footnote BPSO can be implemented using traditional or geometric approaches.

Answer 118

As probability ## Footnote The update rule remains the same, but positions are updated based on this probability.

Answer 119

No explicit velocity; position is updated in discrete Hamming space ## Footnote This approach behaves similarly to multi-parent crossover in genetic algorithms.

Answer 120

* Based on swarm behaviour (no leaders) * New solutions created using velocity update rule * Intuitive parameters (c1, c2, swarm size) * Can be extended to binary problems, dynamic optimisation, discrete search spaces ## Footnote PSO harnesses collective emergent intelligence from simple agents.

Answer 121

FALSE ## Footnote PSO uses personal best and global best memory, while ACO relies on pheromone trails.

Answer 122

* PSO = birds searching for food → everyone moves towards the best smell * Velocity = direction + memory (own best + swarm best) ## Footnote These anchors help in understanding PSO concepts intuitively.

Answer 123

* Maximise power output * Minimise cost ## Footnote These goals conflict, as a more expensive layout may yield more energy, while a cheaper layout may result in lower energy output.

Answer 124

Each solution has M fitness values ## Footnote EAs must modify selection, objective handling, and visualisation to deal with multiple objectives.

Answer 125

Solution a dominates b if: * a is at least as good as b on all objectives * a is better on at least one objective ## Footnote This concept helps in comparing solutions based on multiple objectives.

Answer 126

Set of all non-dominated solutions ## Footnote The front shape indicates the trade-off curve between conflicting objectives.

Answer 127

* Even coverage * Wide spread ## Footnote Bad characteristics include big gaps and clustering in only part of the front.

Answer 128

Use Non-dominated Sorting: 1. Find all Rank-0 individuals (non-dominated) 2. Remove them 3. Find next non-dominated group → Rank-1 4. Continue until all ranked ## Footnote This creates a hierarchy of solution sets.

Answer 129

Selection of non-dominated solutions ## Footnote A new selection operator is needed to decide between non-dominated solutions.

Answer 130

Basic method: 1. Pick two individuals at random 2. If one dominates the other → choose it 3. If neither dominates → choose randomly ## Footnote Improved method introduces a comparison set to increase selection pressure.

Answer 131

Ensure good spread of solutions ## Footnote When two solutions are non-dominated, the one in a less crowded region is preferred.

Answer 132

Divide the solution space into niches ## Footnote Prefer solutions in less populated niches to maintain diversity on the Pareto front.

Answer 133

Non-Dominated Sorting Genetic Algorithm II ## Footnote Key features include elitism, fast non-dominated sorting, and crowding distance for tie-breaking.

Answer 134

Calculates how 'squeezed' a solution is: * Boundary solutions get infinite crowding distance * Others get distance = normalised difference between neighbours ## Footnote Higher distance indicates a more isolated solution, which is preferred.

Answer 135

Steps: 1. Create population (size N) 2. Generate offspring → size 2N 3. Apply fast non-dominated sort 4. Select best ranks until reaching N 5. Sort by crowding distance if last front doesn’t fit 6. Repeat ## Footnote This process ensures the maintenance of diversity and quality in solutions.

Answer 136

Graph interpretation: * Right/top = best power but most expensive * Bottom/left = cheapest but worst power ## Footnote NSGA-II generates a smooth Pareto front of trade-offs.

Answer 137

* No extra parameters * Faster convergence * Maintain best solutions ## Footnote However, they may lead to premature convergence on some problems.

Answer 138

Single-objective EA by weighting objectives ## Footnote Problems include needing multiple runs for different weights and inability to capture non-convex regions of the Pareto front.

Answer 139

* Hard to visualise high-dimensional fronts * Dominance loses discriminative power * Number of solutions needed grows exponentially ## Footnote These challenges complicate the optimisation process.

Answer 140

* Minimise turbines * Minimise area * Minimise annual energy prediction error * Minimise levelized cost of energy * Maximise annual energy output * Maximise efficiency ## Footnote Solved using NSGA-III and visualised using a parallel coordinate plot.

Answer 141

* Used for multiple conflicting objectives * Use Pareto dominance for comparison * Goal: approximate the Pareto front with good diversity * NSGA-II is the most famous and widely used ## Footnote Many-objective optimisation adds significant challenges.

Answer 142

Multiple goals, no single 'best' ## Footnote This helps in remembering the nature of multi-objective optimisation.

Answer 143

Particle Swarm Optimization ## Footnote PSO is a computational method used for optimizing a problem by iteratively improving candidate solutions.

Answer 144

* Position update * Velocity update ## Footnote These updates are based on the best positions found by individual particles and the swarm.

Answer 145

An archive ## Footnote The archive stores non-dominated solutions instead of a single global best solution.

Answer 146

* If current solution dominates pbest → update * If pbest dominates current → keep old one * If neither dominates → two options exist: keep oldest or newest ## Footnote This update mechanism allows for better exploration and exploitation of solutions.

Answer 147

* Stores non-dominated solutions * Represents the current Pareto front * Must be updated with new solutions ## Footnote The archive helps in maintaining a diverse set of solutions for decision-making.

Answer 148

* Large archive slows finding leaders * Too many solutions can be unhelpful for decision-makers ## Footnote Managing archive size is crucial for maintaining efficiency in the optimization process.

Answer 149

* Clustering * Crowding Distance ## Footnote These mechanisms help maintain diversity within the archive by reducing redundancy.

Answer 150

* Use hierarchical clustering on archive solutions * Merge clusters until required number remains * Keep the solution closest to cluster centre ## Footnote This method helps in reducing the number of solutions while retaining representative points.

Answer 151

* Scores solutions by how crowded their region is * Isolated solutions → high score * Crowded solutions → low score ## Footnote This scoring encourages exploration of less populated areas in the solution space.

Answer 152

* Random Leader Selection * Dominated Solution Count * Hypervolume Contribution ## Footnote These methods ensure that leaders are chosen based on their effectiveness and diversity.

Answer 153

Area/volume between solution and a reference point it dominates ## Footnote The leader is the archive solution with the maximum hypervolume contribution.

Answer 154

* Hard to visualize high-dimensional fronts * Reduction in search capability * Exponential growth in needed archive solutions ## Footnote These challenges complicate the optimization process when dealing with four or more objectives.

Answer 155

* Rank archive solutions separately for each objective * Compute average rank ## Footnote Lower average ranks indicate better solutions, facilitating comparison among multiple objectives.

Answer 156

Exploration of sparse regions ## Footnote Solutions in less crowded areas are preferred, promoting diversity in the search space.

Answer 157

FALSE ## Footnote Dominance becomes less effective in many-objective cases, necessitating alternative ranking methods.

Answer 158

* PSO can be adapted to multi-objective problems * Dominance is central but less effective in many-objective cases * Archive stores non-dominated solutions ## Footnote Mechanisms are needed to restrict archive size, maintain diversity, and select leaders.

Answer 159

* Huge calculations instantly * Memory tasks ## Footnote Computers excel at tasks that require speed and accuracy.

Answer 160

* Perception tasks * Learning and generalisation ## Footnote Humans are better at tasks like recognizing faces and finding objects.

Answer 161

* Very fast serial processors * Almost no mistakes * Separate memory and processing * Store information indefinitely ## Footnote These features highlight the efficiency of computers in processing data.

Answer 162

* Massively parallel brain architecture * Can learn, generalise * Processing + memory intertwined * Forget unimportant info * Gracefully degrade ## Footnote These traits allow humans to adapt and function even with damage.

Answer 163

≈ 100 billion ## Footnote This number is comparable to raindrops needed to fill an Olympic swimming pool.

Answer 164

* Dendrites * Cell body * Axon ## Footnote These parts are essential for signal processing and memory in neurons.

Answer 165

* Symbolic AI: Explicit symbols & rules * Connectionism: Implicit numeric representations ## Footnote Symbolic AI is good for reasoning, while Connectionism excels in perception.

Answer 166

Learning occurs through adjusting strength of connections ## Footnote Synapses can strengthen (excite) or weaken (inhibit) connections.

Answer 167

Rumelhart & McClelland (1986) ## Footnote MLPs allowed networks to solve problems like XOR, which are not linearly separable.

Answer 168

* Generalisation * Graceful degradation ## Footnote Generalisation allows NNs to handle new data, while graceful degradation means performance reduces but does not fail completely.

Answer 169

* Rainfall * Air temperature ## Footnote These inputs help predict sewer flow in urban areas.

Answer 170

Mean Absolute Error (MAE) ## Footnote This measure evaluates the accuracy of the model's predictions.

Answer 171

Best performance ## Footnote NSE values range from 1 (best) to -∞ (worst).

Answer 172

perception, learning ## Footnote These tasks often require human-like capabilities.

Answer 173

Cognition via networks ## Footnote This approach mimics the structure and function of the brain.

Answer 174

* Process high-dimensional sensor data * React with low latency * Use little energy * Adapt to dynamic, uncertain environments ## Footnote Traditional ANNs struggle with these requirements.

Answer 175

* More biologically realistic * Much more energy-efficient * Great for time-varying / sequential data * Used in robotics, speech recognition, sensory processing ## Footnote SNNs send information as spikes over time.

Answer 176

* Rate coding * Temporal coding * Population coding ## Footnote Each method has a different approach to representing information through spikes.

Answer 177

A sequence of spikes over time ## Footnote Each neuron receives spikes, accumulates charge, and fires when a threshold is reached.

Answer 178

* Inputs drip in → water level rises * When full → bucket spills (fires a spike) * Water slowly leaks out if unused ## Footnote This creates time-dependent behaviour essential for temporal learning.

Answer 179

* Potentiation: pre-synaptic fires before post-synaptic → connection strengthens * Depression: pre-synaptic fires after post-synaptic → connection weakens ## Footnote STDP is a form of unsupervised learning that matches biological processes.

Answer 180

Spiking function derivative = Dirac delta function: * Zero everywhere * Infinite at 0 * Not differentiable ## Footnote This makes standard backprop impossible for true spikes.

Answer 181

To train SNNs with backprop-style approaches ## Footnote They replace the spike with a smooth approximation, allowing gradient flow.

Answer 182

* Network structure * Weights * Delays * Thresholds * Encoding schemes ## Footnote EAs can evolve multiple components of SNNs.

Answer 183

* Very expensive * Co-evolution complexity ## Footnote SNN evaluation is more costly than ANN evaluation, and multiple components influence each other.

Answer 184

* Biologically realistic, event-driven neural networks * Encode info using rate, timing, or population coding * Neurons spike when charge hits threshold and leak over time * STDP = core unsupervised learning rule * Backprop cannot be used directly (spike derivative is a delta) * Surrogate functions make gradient-based training possible * EAs can optimise SNNs, but expensive * Energy-efficient and ideal for temporal tasks ## Footnote This summary encapsulates the main features and challenges of SNNs.

Answer 185

Hardware inspired directly by the brain, using mixed analogue–digital circuits ## Footnote Modern meaning includes any brain-inspired computing, especially spiking neural networks (SNNs) and specialised processors.

Answer 186

* Memory + processing combined in synapses * Massively parallel * Event-driven * Information stored in spike patterns/timing ## Footnote Neuromorphic systems are more like a living brain than a digital CPU.

Answer 187

* Massive Parallelism * Memory + Processing in One Place * Event-Driven * Scalable ## Footnote These characteristics contribute to its efficiency and speed for specific tasks.

Answer 188

TRUE ## Footnote This is particularly true for tasks like perception, spiking data, or optimisation.

Answer 189

* Routing * Network analysis * Random walks ## Footnote Mapping graph nodes to neurons allows for natural parallelism.

Answer 190

* Integer optimisation * Continuous optimisation * QUBO problems * LASSO regression ## Footnote Spiking systems can solve these problems efficiently due to their energy-based optimisation dynamics.

Answer 191

* Autonomous vehicles * Seizure detection from high-frequency oscillations * Edge computing sensors ## Footnote Spiking is advantageous for time-critical tasks.

Answer 192

* Each chip has 18 ARM cores * 16,000 neurons & 8 million synapses simulated in real time * Uses only 1 watt ## Footnote Designed for large-scale brain simulations & SNN research.

Answer 193

* 1,000,000 neurons * 256 million synapses * Event-driven * Uses the Corelet programming language ## Footnote TrueNorth is designed for brain-like processing.

Answer 194

* Traditional ML outperforms neuromorphic * Training difficulty * Limited accessibility * Lack of benchmarks * Hard programming model ## Footnote These challenges hinder widespread adoption and development.

Answer 195

* Each pixel emits a spike when brightness changes * Results in high temporal resolution * Very low power use ## Footnote They are ideal for robotics, drones, and AR/VR applications.

Answer 196

spike ## Footnote It is much more energy-efficient than classical systems.

Answer 197

* Training difficulty * Usability * Lack of benchmarks ## Footnote These challenges make it still experimental compared to mainstream deep learning.

Nature-Inspired Computing Flashcards

(222 cards)