A while back I posted an exchange between Peter Tse and Neil Levy that focused on parts of Peter's new book, The Neural Basis of Free Will: Criterial Causation. In the wake of that discussion, I asked Peter if he would be interested in writing up an accessible overview of the argument he develops in the book. Fortunately, he was happy to oblige! The following is what he sent me to post here on Flickers. Given the intersection between work in neuroscience and work on the philosophy of action, I think we all need to work a little harder to understand what's happening on the other half of this discplinary divide. In that spirit, I have posted Peter's overview below the fold. Hopefully, everyone will join the ensuing discussion. If he's right, then free will skeptics like me have some work to do!
Absract: In my book I use recent developments in neuroscience to show how volitional mental events can be causal within a physicalist paradigm. (1) I begin by attacking the logic of Jaegwon Kim’s exclusion argument, according to which mental information cannot be causal of physical events. I argue that the exclusion argument falls apart if indeterminism is the case. If I am right, I must still build an account of how mental events are causal in the brain. To that end I take as my foundation (2) a new understanding of the neural code that emphasizes rapid synaptic resetting over the traditional emphasis on neural spiking. (3) Such a neural code is an instance of ‘criterial causation,’ which requires modifying standard interventionist conceptions of causation. A synaptic reweighting neural code provides (4) a physical mechanism that accomplishes downward informational causation, (5) a middle path between determinism and randomness, and (6) a way for mind/brain events to turn out otherwise. This ‘synaptic neural code’ allows a constrained form of randomness parameterized by information realized in and set in synaptic weights, which in turn allows physical/informational criteria to be met in multiple possible ways when combined with an account of how randomness in the synapse is amplified to the level of randomness in spike timing. This new view of the neural code also provides (7) a way out of self-causation arguments against the possibility of mental causation. It leads to (8) an emphasis on deliberation and voluntary attentional manipulation as the core of volitional mental causation rather than, say, the correlates of unconscious premotor computations seen in Libet’s readiness potentials. And this new view of the neural code leads to (9) a new theory of the neural correlates of qualia as the ‘precompiled’ informational format that can be manipulated by voluntary attention, which gives qualia a causal role within a physicalist paradigm. I elaborate each of these ideas in turn below.
(1) Countering Kim’s exclusion argument. The exclusion argument is, roughly, that the physical substrate does all the causal work that the supervenient mental state is supposed to do, so mental or informational events can play no causal role in material events. On Kim’s reductionistic view, all causation seeps away to the rootmost physical level, i.e. particles or strings. Add to that an assumption of determinism, and the laws of physics applicable at the rootmost level are sufficient to account for event outcomes at that level and every level that might supervene on that level. So informational causation, including voluntary mental causation or any type of free will that relies on it, is ruled out.
I argue that indeterminism undermines this sufficiency, so provides an opening whereby physically realized mental events could be downwardly causal. I argue that biological systems introduced a new kind of physical causation into the universe, one based upon triggering physical actions in response to detected spatiotemporal patterns in energy. This is a very different kind of causation than traditional Newtonian conceptions of the causal attributes of energy, such as mass, momentum, frequency or position, which seem to underlie deterministic and exclusionary intuitions. But patterns, unlike amounts of energy, lack mass and momentum and can be created and destroyed. They only become causal if there are physical detectors that respond to some pattern in energetic inputs. Basing causal chains upon successions of detected patterns in energy, rather than the transfer of energy among particles, opens the door not only to informational downward causation but to causal chains (such as mental causal chains or causal chains that might underlie a game of baseball or bridge) that are not describable by or solely explainable by the laws of physics applicable at the rootmost level. Yes, a succession of patterns must be realized in a physical causal chain that is consistent with the laws of physics, but many other possible causal chains that are also consistent with physical laws are ruled out by informational criteria imposed on indeterministic particle outcomes. Physical/informational criteria set in synaptic weights effectively sculpt informational causal chains out of the ‘substrate’ of possible physical causal chains.
(2) A new view of the neural code: I develop a new understanding the neural code that emphasizes rapid and dynamic synaptic weight resetting over neural firing as the core engine of information processing in the brain. The neural code is not solely a spike code, but a code where information is transmitted and transformed by flexibly and temporarily changing synaptic weights on a millisecond timescale. One metaphor is the rapid reshaping of the mouth (analogous to rapid, temporary synaptic weight resetting) that must take place just before vibrating air (analogous to spike trains) passes through, if information is to be realized and communicated. What rapid synaptic resetting allows is a moment by moment changing of the physical and informational parameters or criteria that have to be met before a neuron will fire. This dictates what information neurons will be responsive to and what they will ‘say’ to one another from moment to moment.
(3) Rethinking interventionist models of causation: Standard interventionist models of causation manipulate A to determine what effects, if any, there might be on B and other variables. If instead of manipulating A's output, we manipulate the criteria, parameters or conditions that B places on A's input, which must be satisfied before B changes or acts, then changes in B do not follow passively from changes in A as they would if A and B were billiard balls. Inputs from A can be identical, but in one case B changes in response to A, and in another it does not. This constant reparameterization of B is what neurons do when they change each other's synaptic weights. What I call "criterial causation" emphasizes that what can vary is either outputs from A to other nodes, or how inputs from A are decoded by receiving nodes. On this view, standard interventionist and Newtonian models of causation are a special case where B places no conditions on input from A. But the brain, if anything, emphasizes causation via reparameterization of B, by, for example, rapidly changing synaptic weights on post-synaptic neurons.
(4) How downward causation works: Downward causation means that events at a supervening level can influence outcomes at the rootmost level. In this context it would mean that information could influence particle paths. While it would be impossible self-causation if a supervening event changed its own present physical basis, it is not impossible that supervening events, such as mental information, could bias future particle paths. How might this work in the brain? The key pattern in the brain to which neurons respond is temporal coincidence. A neuron will only fire if it receives a certain number of coincident inputs from other neurons. Criterial causation occurs where physical criteria imposed by synaptic weights on coincident inputs in turn realize informational criteria for firing. This permits information to be downwardly causal regarding which indeterministic events at the rootmost level will be realized; Only those rootmost physical causal chains that meet physically realized informational criteria can drive a postsynaptic neuron to fire, and thus become causal at the level of information processing. Typically the only thing that the set of all possible rootmost physical causal chains that meet those criteria have in common is that they meet the informational criteria set. To try to cut information out of the causal picture here is a mistake; The only way to understand why it is that just this subset of possible physical causal chains—namely those that are also informational causal chains—can occur, is to understand that it is informational criteria that dictate that class of possible outcomes.
The information that will be realized when a neuron’s criteria for firing have been met is already implicit in the set of synaptic weights that impose physical criteria for firing that in turn realize informational criteria for firing. That is, the information is already implicit in these weights before any inputs arrive, just as what sound your mouth will make is implicit in its shape before vibrating air is passed through. Assuming indeterminism, many combinations of possible particle paths can satisfy given physical criteria, and many more cannot. The subset that can satisfy the physical criteria needed to make a neuron fire is also the subset that can satisfy the informational criteria for firing (such as ‘is a face’) that those synaptic weights realize. So sets of possible paths that are open to indeterministic elementary particles which do not also realize an informational causal chain are in essence “deselected” by synaptic settings by virtue of the failure of those sets of paths to meet physical/informational criteria for the release of a spike.
(5) Between determinism and randomness: Hume (1739) wrote “’tis impossible to admit of any medium betwixt chance and an absolute necessity.” Many other philosophers have seen no middle path to free will between the equally ‘unfree’ extremes of determinism and randomness. They have either concluded that free will does not exist, or tried to argue that a weak version of free will, namely, ‘freedom from coercion,’ is compatible with determinism.
A strong conception of free will, however, is not compatible with either determined or random choices, because in the determined case there are no alternative outcomes and things cannot turn out otherwise, while in the random case what happens does not happen because it was willed. A strong free will requires meeting some high demands: Beings with free will (a) must have information processing circuits that have multiple courses of physical or mental activity open to them; (b) they must really be able to choose among them; (c) they must be or must have been able to have chosen otherwise once they have chosen; and (d) the choice must not be dictated by randomness alone, but by the informational parameters realized in those circuits. This is a tough bill to fill, since it seems to require that acts of free will involve acts of self-causation.
Criterial causation offers a middle path between the two extremes of determinism and randomness that Hume was not in a position to see, namely, that physically realized informational criteria parameterize what class of neural activity can be causal of subsequent neural events. The information that meets preset physical/informational criteria may be random to a degree, but it must meet those criteria if it is to lead to neural firing, so is not utterly random. Preceding brain activity specifies the range of possible random outcomes to include only those that meet preset informational criteria for firing.
(6) How brain/mind events can turn out otherwise: The key mechanism, I argue, whereby atomic level indeterminism has its effects on macroscopic neural behavior is that it introduces randomness in spike timing. There is no need for bizarre notions such as consciousness collapsing wave packets or any other strange quantum effects beyond this. For example, quantum level noise expressed at the level of individual atoms, such as single magnesium atoms that block NMDA receptors, is amplified to the level of randomness and near chaos in neural and neural circuit spiking behavior. A single photon can even trigger neural firing in a stunning example of amplification from the quantum to macroscopic domains. The brain evolved to harness such ‘noise’ for information processing ends. Since the system is organized around coincidence detection, where spike coincidences (simultaneous arrival of spikes) are key triggers of informational realization (i.e. making neurons fire that are tuned to particular informational criteria), randomizing which incoming spike coincidences might meet a neuron's criteria for firing means informational parameters can be met in multiple ways just by chance.
(7) Skirting self-causation: A synaptic account of the neural code also gets around some thorny problems of self-causation that have been used to argue against the possibility of mental causation. The traditional argument is that a mental event realized in neural event x cannot change x because this would entail impossible self-causation. Criterial causation gets around this by granting that present self-causation is impossible. But it allows neurons to alter the physical realization of possible future mental events in a way that escapes the problem of self-causation of the mental upon the physical. Mental causation is crucially about setting synaptic weights. These serve as the physical grounds for the informational parameters that must be met by unpredictable future mental events.
(8) Voluntary attention and free will: I argue that the core circuits underlying free choice involve frontoparietal circuits that facilitate deliberation among options that are represented and manipulated in executive working memory areas. Playing out scenarios internally as virtual experience allows a superthreshold option to be chosen before specific motoric actions are planned. The chosen option can best meet criteria held in working memory, constrained by conditions of various evaluative circuits, including reward, emotional and cognitive circuits. This process also harnesses synaptic and ultimately atomic level randomness to foster the generation of novel and unforeseeable satisfactions of those criteria. Once criteria are met, executive circuits can alter synaptic weights on other circuits that will implement a planned operation or action.
(9) A new theory of qualia: The paradigmatic case of volitional mental control of behavior is voluntary attentional manipulation of representations in working memory such as the voluntary attentional tracking of one or a few objects among numerous otherwise identical objects. If there is a flock of indistinguishable birds, there is nothing about any individual bird that makes it more salient. But with volitional attention, any bird can be marked and kept track of. This salience is not driven by anything in the stimulus. It is voluntarily imposed on bottom-up information, and can lead to eventual motoric acts, such as shooting or pointing at the tracked bird. This leads to viewing the neural basis of attention and consciousness as not only realized in part in rapid synaptic reweighting, but also in particular patterns of spikes that serve as higher level units that traverse neural circuits and open what I call the ‘NMDA channel of communication.’ Qualia are necessary for volitional mental causation because they are the only informational format available to volitional attentional operations. Actions that follow volitional attentional operations, such as volitional tracking, cannot happen without consciousness. Qualia on this account are a ‘precompiled’ informational format made available to attentional selection and operations by earlier, unconscious information processing.
Conclusion: Assuming indeterminism, it is possible to be a physicalist who adheres to a strong conception of free will. On this view, mental and brain events really can turn out otherwise, yet are not utterly random. Prior neuronally realized information parameterizes what subsequent neuronally realized informational states will pass presently set physical/informational criteria for firing. This does not mean that we are utterly free to choose what we want to want. Some wants and criteria are innate, such as what smells good or bad. However, given a set of such innate parameters, the brain can generate and play out options, then select an option that adequately meets criteria, or generate further options. This process is closely tied to voluntary attentional manipulation in working memory, more commonly thought of as deliberation or imagination. Imagination is where the action is in free will.