Conditioning in mammals was first described by Fetz and Baker (1973) who trained monkeys using juice reward to modifying the firing rates of single units (SUs). Monkeys succeeded in increasing the firing rate of one SU and simultaneously decreasing the firing rate of a separate SU recorded through the same electrode. In more recent years, monkeys were trained to increase or decrease firing rates of single units by operant conditioning, sometimes over several days (Moritz and Fetz, 2011). This important result was taken a step further by developing a bi-directional brain computer interface, whereby the conditioning was promoted by electrical stimulation of neurons in surrounding cortex (Fetz, 2015). A “Neurochip” was developed that provides recording and stimulation in a closed loop system that enhances conditioning (Shupe et al., 2021). Additional studies in monkeys report increased firing using operant conditioning and cortical stimulation via the Neurochip (Eaton et al., 2017).

Operant conditioning of SUs in animals has been performed in hippocampus in rabbits (Berger et al., 1983), visual cortex in cats (Shinkman et al., 1974), motor cortex in rats (Arduin et al., 2014), cerebellum in rabbits (Katz and Steinmetz, 1997), parietal cortex in monkeys (Steege et al., 1982), H-reflex in rats (Carp et al., 2001) and globus pallidus in rabbits (Richardson and Thompson, 1985), among other reports. Operant conditioning in human studies have been restricted to EEG and its various bandwidths (Rockstroh et al., 1984). The H-reflex has also been conditioned in humans (Shindo et al., 1994). As the above brief review illustrates, however, conditioning of available SUs in animals or humans can increase or decrease firing rates of SUs, depending on the task. Until now, there are no known reports of conditioning the interface between human single unit firings and computers or machines.

The question addressed in this report seeks to understand the SU firing mechanisms that underlie conditioning in humans. During a decade long study in participant 5 (FDA IDE G960032S), there was an opportunity to study conditioning in a silent singing task. The participant was presented with an audible tone for 20 s, then attempted to silently sing the tone in his head for 10 s, with a prior resting control period for 10 s. Twenty single units were recorded simultaneously while the tone was attached to one single unit, so that he received feedback of the tone every time the SU fired. We theorized that during this task (a) some of the inaudible SUs would be recruited during conditioning, (b) that the recruitment would be task specific, and (c) that different tones would recruit different patterns of SU activity. The data presented here supports these theoretical conjectures.

The importance of these findings plays directly into attempts to provide a high bandwidth interface especially in the face of few recorded SUs. The current emphasis in research on brain-computer interfaces is the simultaneous recording of many hundreds, if not thousands, of neural firings of multi-units (Schwartz, 2004; Ajiboye et al., 2017; Musk and Neuralink, 2019). With conditioning, and the underlying recruitment of more SUs as reported here, the number of SUs can be more modest, though the actual number required in a specific task is still undefined.

The firing rates of SUs vary widely as shown in Figure 6. Because the neuropil grows in both ends of the glass tip, it is not surprising that Betz cells from layer five and interneurons from layer 2 or 3 would grow into the electrode tip from above (interneurons) and below (Betz). Because interneurons usually have fast firing rates and Betz cells usually have lower firing rates, it is not surprising to see the distribution as shown in Figure 6.

The results above indicate that some of the non-audible single units were recruited during conditioning as evidenced by the symmetrical firing rates during the three phases (listen, rest, and silent sing) when tone feedback, especially volume feedback, was provided. Importantly, the conditioning improved from one trial to the next as shown in Figure 7. Also, overall firing rates increased as the trials progressed from no feedback to feedback, such that some SUs had a dramatic increase in firing rates after the tone feedback was introduced as can be seen in Figure 7 for both tones. Comparing tone 262 Hz and tone 523 Hz conditioning, there is a sustained “flipping” of firing rate patterns in SU 2-17 as can be seen by the “V” shape of the plotted firing rates. More evidence for identifying the tone from these SU firing patterns is discussed below.

Some SUs were recruited differentially depending on the phase of the paradigm (listen, rest, or silent sing) as shown in Figure 8. Only three examples are shown, but a power point slide in Supplementary Material will show more examples. The lines indicate which SUs had enhanced (green) or inhibited firing (red or brown).

Single unit firing patterns differed between different tone frequencies, 256 versus 523 Hz, as illustrated in Figure 9. It is apparent that the tone could be recognized from the pattern of single unit firing patterns in these examples. In addition, the patterns were specific for the phase of the task (listen, rest, and silent sing), making it likely that the phase could be recognized. Pattern recognition is very important because it is the basis of detecting not just tones, but other modalities such as speech.

No other researchers are known to have published conditioning of human SU data. Instead, other research paradigms use multi-unit activity as a proxy for SU activity. To test the viability of this proxy, we extracted the multi-units from the data along with the SUs on day 1556. Figure 10 illustrates that there is no symmetry to the firing activity of multi-units when compared to the SU data. In other words, the precision available with SU firings is not available with multi-unit activity. Thus this proxy is not valid.

From a practical point of view, multi-units may be adequate for neural prostheses and are being used as such (Hochberg et al., 2012; Homer et al., 2013; Bouton et al., 2016; Ajiboye et al., 2017; Musk and Neuralink, 2019). However, as mentioned above, high bandwidth data will improve precision decoding, and thus improve the quality of speech (or fine finger movement) decoding.

Other workers have used single units with interesting results. Ossmy et al. (2015) recorded human auditory cortex during surgery and detected SU activity related to onset of speech. Studies comparing audible speech and hand use in American Sign Language were studied: Single unit recordings in the anterior temporal lobe, at a site later identified to be important in handshape formation, showed sustained activity during naming with superimposed increases in activity during audible speech (Haglund et al., 1993). No evidence is presented indicating decoding for speech using SUs in these studies.

Other researchers use the frequency domain to decode speech with interesting results. Ramsey’s group use high density ECoG electrode grids to successfully decode phonemes by placing it over the frontal lobe and recording neural activity in the frequency domain (Ramsey et al., 2017). Chang’s group place a grid of ECoG electrodes over the speech areas and reproduce speech using the frequency domain (Conant et al., 2018; Hamilton et al., 2018). More recently, Moses et al. (2021) decoded speech in a locked-in patient using high density ECoG electrodes in real time using data acquired from the gamma band. Decoding was augmented by a natural language model and a Viterbi decoder. Results indicate decoding at 15.2 words per minute with a median error rate of 25.6% (Moses et al., 2021). All these efforts using ECoG recordings are ongoing and may result in natural speech from locked-in people.

Using ECoG recordings, Chang and his group have identified the human lateral primary cortex as the area for control of articulation, in other words, the speech motor cortex (Bouchard et al., 2013). In our studies, this area has been confirmed by recording SUs and detecting phones, words and phrases Brumberg et al. (2011). Hence, implanting the motor speech cortex to restore speech in those locked-in people who have an intact cortex is essentially a motor, not a speech, decoding task (Kennedy et al., 2011).

The present data indicate that SUs can be conditioned to the task. Though the firing rates were initially random, they began to fire at similar rates as audible feedback was introduced. It is important that the feedback was audible as conditioning occurred only when one SU fired while the other 19 SUs became conditioned as the trials progressed. Thus even though the SUs were not conditioned to the task initially, the SUs became conditioned with feedback. Conditioning with feedback is an important capability of SUs because they expand the bandwidth of the population of recorded neurons and thus enhance the brain to machine/computer interface. It is important to recall that multi-units are not a proxy for SUs in this task but they are used successfully in brain to computer interfaces. The pattern of SU firing is one of the key aspects of decoding and forms the basis of speech decoding (Brumberg et al., 2011).

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by the Neural Signals, Inc., Ethics Committee. The patients/participants provided their written informed consent to participate in this study.

PK performed the experiment and analyzed the data. AC performed the surgeries. Both authors contributed to the article and approved the submitted version.

PK has 98% ownership of Neural Signals, Inc. The remaining author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.