Science may mask what it tries to explain

Receptive fields (RF) are a research area whose development can be potentially affected by a scientific methodology that is based on making distinctions. In this particular arena, scientists would like to explain how receptive fields respond to visual stimuli, and how the underlying circuity in the visual cortex is able to decode such responses and create an effective internal representation of the environment. After Hubel (1981) demonstrated that the responses of cortical cells are associated with properties of the visual image, a great paradigm shift took place. Scientific research became guided by an assumption stated by Hubel himself: “understanding how [receptive fields] respond to some visual stimuli and ignore others allow us to predict how a cell will react to any given visual scene.” Thus, the resulting methodology was concerned with how a certain RF is different from another, but most importantly, how responses within a single receptive field are distinct depending on the varying stimuli and its context. Interestingly, this process of science may potentially create a discontinuity in trying to explain the role of receptive fields in the processing of visual information.

    In particular, experiments with highly restricted visual stimuli can lead to some insight about the inner workings of RF but there is no guarantee that such studies will generalize to more complex situations (Olshausen & Field, 2005).  First, the visual system was evolutionarily designed to respond to changes in our environment—consisting of natural scenes— and not in artificial conditions like those studied extensively. Second, the visual system is in a “continuing process of normalization and calibration” (Gilbert, 1996).

Thus, without having much insight about how such process occurs, and the amount of temporal information that the visual system needs to code, we cannot predict the effect of previous visual stimulation, or even the effect of experimental methodology. For instance, it has been observed that substantial changes in size and position of the RF can be induced by previous visual stimulation and that “whatever one does to measure the response properties of a cell may change them” (Gilbert, 1996).  Thirdly, there is a high level of non-linearity inherent in neural computations in the visual system (excluding complex processes like lateral inhibition), which strongly suggests that categorizing stimuli as well as cell responses may in fact be counterproductive for furthering our understanding. This nonlinearity is easily observed in artificial neural network, and our inability to understand how this simplified version of the biological neuron is creating sparse and effective internal representation (except under very constricted set of examples).

    Using this methodology, there is also an inherent process of distinguishing which cells to study. As Olshausen & Field (2005) argue, there is a bias to prefer to study neurons that are “visually responsive,” with large cell bodies and extracellular action potentials, and also neurons with high firing rates. A final aspect that can mask the continuity of what we are trying to explain can take place because of pressures from journals to make data look tidy (Olshausen & Field, 2005), which may result in limited theories that oversimplify, and bias towards theories that seem to model observations, and not those that fail to do so.
It is quite possible that a process of making distinction could eventually explain the mysteries of receptive fields. Given the complexity of the system, I believe that we will continue using this process of science until more complex computational models are developed, models whose nonlinearity and adaptability we can understand. In particular, we will be able to fill the gap of how RF respond in varying situations, from simple lines in Hubel’s original experiments to highly complex natural scenes.