Abstract
Neuroscience has focused on the detailed implementation of computation, studying neural codes, dynamics and circuits. In machine learning, however, artificial neural networks tend to eschew precisely designed codes, dynamics or circuits in favor of brute force optimization of a cost function, often using simple and relatively uniform initial architectures. Two recent developments have emerged within machine learning that create an opportunity to connect these seemingly divergent perspectives. First, structured architectures are used, including dedicated systems for attention, recursion and various forms of short- and long-term memory storage. Second, cost functions and training procedures have become more complex and are varied across layers and over time. Here we think about the brain in terms of these ideas. We hypothesize that (1) the brain optimizes cost functions, (2) these cost functions are diverse and differ across brain locations and over development, and (3) optimization operates within a pre-structured architecture matched to the computational problems posed by behavior. Such a heterogeneously optimized system, enabled by a series of interacting cost functions, serves to make learning data-efficient and precisely targeted to the needs of the organism. We suggest directions by which neuroscience could seek to refine and test these hypotheses.
Introduction
Historically, neuroscience and machine learning (ML) have operated under different paradigms. Neuroscience has traditionally focused on the "how"—the intricate details of neural codes, cell types, and circuits. Conversely, machine learning has focused on the "what"—the optimization of cost functions to achieve specific behaviors, often utilizing relatively uniform architectures.
However, this divergence is narrowing. Modern ML is increasingly adopting complex, structured architectures and varied training schedules. Meanwhile, neuroscience faces the challenge of interpreting massive datasets without a functional "top-down" framework. This paper proposes that the brain functions as a sophisticated optimization machine, bridging these fields through a framework of internally generated cost functions and specialized modules.
The Three Core Hypotheses
1. The Brain Optimizes Cost Functions
The first hypothesis posits that the brain utilizes mechanisms to optimize cost functions, analogous to backpropagation in artificial neural networks. While the exact biological implementation remains a subject of research, candidates include Feedback Alignment, Spike-Timing-Dependent Plasticity (STDP), and dendritic computation.
2. Diverse and Evolving Cost Functions
Unlike simple ML models that use a single global error signal, the brain likely employs a diverse array of cost functions that change over development. This "bootstrapping" process allows simple, innate signals to train more complex systems. For instance, an innate reflex to track faces can provide the training data for a sophisticated visual recognition system.
3. Optimization within Specialized Architectures
Optimization does not occur in a vacuum but within a pre-structured architecture. ML concepts such as content-addressable memory, buffers, and gated relays find biological parallels in the hippocampus, prefrontal cortex, and thalamus.
Key Insights: The Society of Optimized Networks
The brain is viewed as a "Society of Mind," where specialized systems each optimize their own cost functions. This modularity makes learning significantly more data-efficient than standard end-to-end deep learning.
One critical concept is Cost Function Bootstrapping. Evolution may not hard-wire complex behaviors directly but instead hard-wires the cost functions that necessitate learning those behaviors. Furthermore, the brain may spatialize time; memory systems like the hippocampus store sequences, allowing the brain to optimize over a "map" of the past, circumventing the biological implausibility of unfolding networks through time.
The cortical sheet can be viewed as a general-purpose optimization engine, while subcortical structures like the basal ganglia act as cost function generators, directing the cortex on what to learn based on environmental and internal needs.
Future Directions and Predictions
This framework leads to several testable predictions:
- Gradient Search: Neural weight changes should be observable as moving "down the gradient" of specific objectives.
- BMI as a Tool: Brain-Machine Interfaces can be used to impose artificial cost functions, testing the adaptability and optimization capabilities of neural circuits.
- Functional Mapping: Researchers should look for molecular and structural signatures of different cost functions across brain regions.
Limitations
Identifying cost functions from observed dynamics is inherently difficult, as multiple objectives can produce similar behaviors. Additionally, while biologically plausible versions of backpropagation exist, their actual use in the brain is still being verified. Finally, the traditional focus on a universal "neural code" may be misleading if such codes are merely byproducts of area-specific optimization.
Conclusion
By shifting focus from specific circuits to the objectives they satisfy, we can begin to reverse-engineer the brain's remarkable efficiency. Integrating structural mapping, large-scale recording, and new mathematical tools for gradient identification will be essential for testing this unified theory of brain function.
Summary updated