ImageNet Competition and Deep Learning

ImageNet Competition and Computer Vision

• ImageNet competition spurred advancements in computer vision.
• Models trained on a large dataset and evaluated on a separate test set.
• Top-5 error rate metric measures accuracy.
• Early models showed small improvements.
• AlexNet (2012) using deep neural networks achieved strong results.
• Deeper networks (2014) significantly reduced error rates triggering rapid use of deep learning.

Deep Neural Networks (DNNs)

• DNNs are "deep" due to multiple layers enabling reuse of expressions.
• Layers extract increasingly complex features.
• Early layers detect simple features (edges, textures).
• Deeper layers detect more complex features (shapes, objects).
• Richer representations generalize better.

Convolutional Neural Networks (ConvNets)

• ConvNets overcome traditional neural networks' limitations.
• Traditional neural nets have only one hidden layer.
• ConvNets use multiple hidden layers for better generalization.
• Gradual transformation of input to output via hierarchical feature extraction.
• Improved generalization ability.

Convolutional Layer Properties

• Receptive field size: Input patch size processed by one neuron (width x height pixels).
• Depth: Number of filters (each learns a specific feature).
• Activation map: Output of input image after applying a kernel (filter).
• Weights: Filters have weights determined by the receptive field size.
• Number of receptive fields: Total neurons (receptive fields) in the feature map.

Local Filters (Kernels)

• Kernels operate on small input regions (receptive fields) to extract features.
• Each kernel learns efficient patterns.

Benefits of Priors in Learning

• ConvNets incorporate locality priors (local patterns).
• Inspiration from visual system (neurons process local patterns).
• Inductive bias reduces computational load.
• Faster and more efficient learning, requiring fewer samples.

Brain vs. Computer Architecture

• Brain: Slow conduction speed, low energy consumption, massively parallel processing
• Computers: Fast transduction, massive energy usage, massively serial processing
• DNNs bridge the gap by integrating memory and processing in weights (analogous to synapses).

GOFAI vs. Statistical Learning

• GOFAI: Symbolic reasoning, explicit rules, excels at planning and logic.
• Statistical Learning: Numerical optimization, (DNNs), excels at pattern recognition, is a black box.

DNNs and Linking Brain Behavior

• Physical shape: low-level details.
• Perceived shape: high-level, abstract representations.
• DNN layers reflect this hierarchical processing.
• Early layers correspond to physical shapes; deeper layers to perceived shapes.
• Mirror human visual processing.

DNN Families

• Cornet family: Mimics ventral visual stream and feedforward processing.
• ResNet family: Utilizes residual connections for very deep networks and mimics recurrent processing.

Feedforward vs. Recurrent Processing in Object Detection

• Feedforward: Rapid object detection, single pass, input to output.
• Recurrent: Feedback loops, crucial for complex tasks (e.g., scenes).
• Backward masking interrupts further processing to study recurrent effects.
• Simple scenes use feedforward processing; complex scenes rely on recurrent processing.

Networks and Behaviour

• ResNet networks align with human behavior in unmasked conditions, and shallow networks in masked ones.
• Cornet align with masked conditions.

DNNs and Human Brain Comparisons

• Similar stimuli (e.g., faces) yield similar activation patterns.
• Dissimilar stimuli (e.g., face vs house) yield distinct activation patterns.
• DNN layers align with levels of brain areas (V1, IT cortex).
• DNNs may reflect hierarchical structure of human visual processing.

Using DNNs in Research

• A framework for studying DNNs (e.g., Measuring brain data, measuring model data, comparing outcomes).
• Use DNNs as models in animal experiments

Differences between Physical and Behavioral Sciences

• Physical: Interpretation and prediction closely aligned, simple formulas describe phenomena.
• Behavioral: Large gap between theory and tasks, interpretation and prediction may not quite match.

Critique of DNNs as Behavioural Models

• DNNs often lack cognitive understanding.
• They simply replace one black box (brain) with another (DNN).
• Cloning a system (with accurate replication) may lead to few insights into actual understanding.

DNNs as Alternative Behavioural Models

• DNNs can predict real-world behavior and can be useful tools for scientists with limitations to study and potentially control complex systems, like predicting behaviors.

How to Use Neural Networks as Models

• How the brain computes (mechanisms, circuits, operations).
• What the brain computes (abstractions, representations, statistics).
• Why the brain computes the way it does (learning, goals, organization).