Professional Documents
Culture Documents
artificial intelligence
tention, have been incorporated into deep
network models, but in general, almost
everything that we know about neurons—
their structure, types, interconnectivity, and
Combining deep learning with brain-like innate structures so on—was left out of deep-net models in
may guide network models toward human-like learning their current form. It is currently unclear
which aspects of the biological circuitry are
computationally essential and could also be
By Shimon Ullman ing input patterns. Successful learning will useful for network-based AI systems, but
W
cause the network to go beyond memoriz- the differences in structure are prominent.
hen the mathemati- ing the training examples, and be able to For example, biological neurons are highly
cian Alan Turing posed generalize, and provide correct outputs to complex and diverse in terms of their mor-
the question “Can ma- new input patterns, which were not en- phology, physiology, and neurochemistry.
chines think?” in the countered during the learning process. The inputs to a typical excitatory pyrami-
first line of his seminal Comparisons of deep network models dal neuron are distributed over complex,
1950 paper that ush- with empirical physiological, functional highly branching basal and apical dendritic
ered in the quest for artificial intelligence magnetic resonance imaging, and behav- trees. Inhibitory cortical neurons come in
(AI) (1), the only known systems carrying ioral data have shown some intriguing simi- a variety of different morphologies, which
out complex computations were biological larities between brains and the new models are likely to perform different functions.
nervous systems. It is not surprising, there- (4), as well as dissimilarities (5) (see the None of this heterogeneity and other com-
fore, that scientists in the nascent field of figure). In comparisons with the primate plexities are included in typical deep-net
AI turned to brain circuits as a source for visual system, similarities between physi- models, which use instead a limited set of
guidance. One path that was taken since ological and model responses were closer highly simplified homogeneous artificial
the early attempts to perform intelligent for the early compared with later parts of neurons. In terms of connectivity between
computation by brain-like circuits (2), and the neuronal responses, suggesting that the units in the network, cortical circuits in the
which led recently to remarkable successes, deep network models may capture better brain are more complex than current deep
can be described as a highly reductionist the early processing stages, compared with network models and include rich lateral
approach to model cortical circuitry. In its later, more cognitive stages. connectivity between neurons in the same
basic current form, known as a “deep net- In addition to deep nets, AI models re- layer, by both local and long-range connec-
work” (or deep net) architecture, this brain- cently incorporated another major aspect tions, as well as top-down connections go-
inspired model is built from successive of brain-like computations: the use of re- ing from high to low levels of the hierarchy
layers of neuron-like elements, connected inforcement learning (RL), where reward of cortical regions, and possibly organized
by adjustable weights, called “synapses,” signals in the brain are used to modify in typical local “canonical circuits.”
after their biological counterparts (3). The behavior. Brain mechanisms involved in The notable successes of deep network–
application of deep nets and related meth- this form of learning have been studied based learning methods, primarily in prob-
ods to AI systems has been transformative. extensively (6), and computational models lems related to real-world perceptual data
They proved superior to previously known (7) have been used in areas of AI, in par- such as vision and speech, have recently been
methods in central areas of AI research, ticular in robotics applications. RL is used followed by increasing efforts to confront
including computer vision, speech recogni- in the context of an agent (a person, ani- problems that are more cognitive in nature.
tion and production, and playing complex mal, or robot) behaving in the world, and For example, in the domain of vision, net-
games. Practical applications are already receiving reward signals in return. The work models were developed initially to deal
in broad use, in areas such as computer vi- goal is to learn an optimal “policy,” which with perceptual problems such as object clas-
sion and speech and text translation, and is a mapping from states to actions, so as to sification and segmentation. Similar meth-
large-scale efforts are under way in many maximize an overall measure of the reward ods, with some extensions, are now being
other areas. Here, I discuss how additional obtained over time. RL methods have been applied to higher-level problems such as im-
aspects of brain circuitry could supply cues combined in recent AI algorithms with deep age captioning, where the task is to produce a
for guiding network models toward broader network methods, applied in particular to short verbal description of an image, or to the
aspects of cognition and general AI. game playing, ranging from popular video domain of visual question answering, where
The key problem in deep nets is learning, games to highly complex games such as the task is to produce adequate answers to
which is the adjustment of the synapses to chess, Go, and shogi. Combining deep nets queries posed in natural language (that is,
produce the desired outputs to their input with RL produced stunning results in game human communication) about the content
patterns. The adjustment is performed au- playing, including convincing defeats of the of an image. Other, nonvisual tasks include
tomatically based on a set of training exam- world’s top Go players, or reaching a world- judging humor, detecting sarcasm, or captur-
ples, which are provided by input patterns champion level in chess after ~4 hours of ing aspects of intuitive physics or social un-
coupled with their desired outputs. The training, starting from just the rules of the derstanding. Similar methods are also being
learning process then adjusts the weights game, and learning from games played in- developed for challenging real-world applica-
to produce the desired outputs to the train- ternally against itself (8). tions such as online translation, flexible per-
From the standpoint of using neurosci- sonal assistants, medical diagnosis, advanced
Department of Computer Science, Weizmann, Institute of ence to guide AI, this success is surpris- robotics, or automatic driving.
Science, Rehovot, Israel. Email: shimon.ullman@weizmann.ac.il ing, given the highly reduced form of the With these large research efforts, and the
leans heavily toward the empiricist side, us- man cognitive system. Recent modeling of
ACK N OW LE D G M E N TS
ing relatively simple and uniform network visual learning in infancy (13) has shown a
Supported by European Union’s Horizon 2020 Framework
structures, and relying primarily on ex- useful combination of learning and innate
785907 (HBP SGA2). S. U. thanks colleagues at the Center
tended learning, using large sets of train- mechanisms, where meaningful complex for Minds, Brains and Machines at Massachusetts Institute of
ing data. By contrast, biological systems concepts are neither innate nor learned on Technology for helpful discussions.
often accomplish complex behavioral tasks their own. The innate components in this 10.1126/science.aau6595