By Andrew Saxe, Stephanie Nelli, Christopher Summerfield. How can neuroscientists use DL to understand biological brains? What are conceptual & methodological challenges of comparing behaviour, learning dynamics, and neural representation in artificial and biological systems? Resurgence of 1980s PDP ideas: current compute can learn from raw data, as opposed to hand-extracted features.
Limitations of the DL framework
If neural computation emerges uncontrollably through blind, unconstrained optimisation, then how can neuroscientists formulate new, empirically testable hypotheses about neural mechanism? Such hypotheses are argued to take the form of design choices about learning rules, regularisation principles, or architectural constraints in deep networks.
Comparing neural and deep representations is statistically challegning. A popular approach is to learn a linear mapping from neurons to network units, and to evaluate the predictive validity on a held-out dataset. DNNs can explain 60% of variance, but DNNs that don’t classify images correctly can explain 55%.
With RSA, it’s unclear what type of agreement is being tested.
Structure in animal learning
Animal behaviour is richly structured, in theory permitting researchers to make systematic comparisons with machine performance. E.g., animal decisions are subject to stereotyped biases, but also irreducibly noisy; animals are flexible but forgetful, behaving as if memory and control systems were capacity limited, and the rate and effectiveness of information acquisition depends strongly on the structure and ordering of the study materials.
They go on to explain progressively differentiated structured learning in children (animal v plant is learned before rose v daisy). Even when they give identical solutions, deep linear networks exhibit similar patterns while shallow ones don’t. Highlight the importance of studying learning dynamics and representational evolution during training.
Artificial vs natural neural
Comparison problem (ANN vs human): how strong are test subjects’ priors? Nature v nurture: are faces innate or acquired?
Deep learning asks how neural codes emerge from different learning principles: traditionally supervised learning, but also suggested others (e.g. Hebbian).
a successful AI model that has yet to impact neuroscience proposes instead that representation formation is driven by the need to accurately predict the motivational value of experience.
Problems arise due to amounts of labeled data required and credit assignment. Grand challenge for neurosci: can learning in the brain assign credit across the neural hierarchy? If so, identify biologically realistic implementation:
- one where updates are local, and
- forward and backward network connectivity is not required to be symmetric
Hope: ML will soon offer more powerful models in which higher cog functions emerge naturally via a “blind search” process. DL has been fused with RL, context-addressable memory, MCTS. But let’s embrace prev neurosci work: we have understanding of some cog systems, e.g. the navigation system in the rodent medial temporal lobe, the motor system in song birds, the saccadic system in the macaque monkey.
Catastrophic interference & resource allocation during learning
Generally, parameterisation that solves task A is not guaranteed to solve any other; training on task B, gradient descent drives network weights away from the local minimum for A.
Evidence grows that offline replay may be important for memory consolidation; continual learning problem begs the question: is biological learning actively partitioned so as to avoid catastrophic interference? They claim “Animals don’t always benefit from interleaved study (e.g. cello+violin),” but I’m not convinced.
More general: how have biological systems evolved to both minimize negative transfer (interference) and maximize positive (generalization) between tasks? Some theories:
the brain has found a solution by promoting shared neural codes, which in turns allows for strong transfer, but deploying control processes to gate out irrelevant tasks that might provoke interference. They suggest that this answers the question of why, despite a brain that comprises billions of neurons and trillions of connections, humans struggle with multi-tasking problems such typing a line of computer code whilst answering a question.