Danko Nikolic | Frankfurt Institute for Advanced Studies Goethe University (original) (raw)
Papers by Danko Nikolic
arXiv (Cornell University), Mar 26, 2023
Handbuch Data Science und KI
Neural activity in vitro can show bursts of activity. These bursts are termed neural avalanches a... more Neural activity in vitro can show bursts of activity. These bursts are termed neural avalanches and their size distribution f(s) approximates a power law [Beggs Plenz, 2003]. Since power law distributions are characteristic for self-organized critical (SOC) states [Bak, Tang, Wiesenfeld, 1987], neural activity was proposed to be SOC, too. Moreover, SOC may provide a basis for optimal information processing [Shew Plenz, 2013]. Evidence for the SOC hypothesis has been obtained for coarse measures of neural activity (LFP, EEG, MEG, BOLD), but surprisingly for spiking activity in vivo evidence for SOC is still missing. Therefore we analyzed highly parallel spike recordings from rats (hippocampus), cats (visual cortex) and monkeys (prefrontal cortex). For all recordings the f(s) were similar (Fig. 1 A-C), but showed fundamental differences to f(s) from a critical spiking model (Fig. 1 D), even under subsampling (Fig. 1 E). The differences between in vivo dynamics and model dynamics could be overcome by decreasing the model’s excitatory synaptic strength, while increasing its external input commensurately (Fig. 1 F). Thereby the model became subcritical and its separation of time scales (STS), which is fundamental to SOC, was eliminated. The match between the subcritical model and the neural activity held for standard and novel avalanche measures (f(s); branching parameter; mean avalanche size; frequency of single events) even when changing the temporal bin size over its full range. In addition, we showed that the same results held for local field potentials recorded in humans. These results suggest that neural activity in vivo is not SOC, but instead reflects a slightly subcritical regime without STS. This regime strikes a balance between optimal information processing and the need to avoid runaway activity. In this regime, avalanches are not temporally separated bursts, but form a mélange. Potential advantages of this regime compared to SOC are faster information processing (due to the lack of STS) and keeping a safety margin from supercriticality, which has been linked to epilepsy
Frontiers in Systems Neuroscience, 2019
Machine learning has achieved remarkable success with deep learning technologies. However, these ... more Machine learning has achieved remarkable success with deep learning technologies. However, these methods are often inefficient in terms of resources; they require large datasets, many parameters and consume much computational power. In this paper, I define a general strategy for machine learning, named strong machine learning, which aims to create resource-effective machine learning models. Under strong machine learning fall all the approaches that learn inductive biases during an initial phase and later apply those inductive biases to make models more effective learners. Several strong machine learning methods already exist and some are very popular exactly due to their effectiveness. However, strong machine learning is in its infancy and a lot more can be done. In order to further advance AI, we need to direct our effort toward developing even better, more powerful strong machine learning methods.
Computational Biology and Chemistry
The Handbook of Data Science and AI
Is clustering necessary? The mean rate classifier is applied directly on the spike-trains and the... more Is clustering necessary? The mean rate classifier is applied directly on the spike-trains and therefore is not affected by clustering. The trajectory classifier can also operate directly on the unclustered activity vectors because it relies just on Euclidean distances. However, the2 specificity classifier is dependent on clustering because it needs a finite set of patterns on which to compute statistics such as specificity and frequency of occurrence. Activity vectors cannot be directly used in this latter case and clustering must be performed to identify model vectors, because of the following reasons. Our approach (and others as well – see Baker and Gerstein [1]) does not binarize and bin multineuronal spike data, as is the case in other methods [2, 3] where a finite set of 2n vectors (n is the number of neurons) is obtained. By convolving spike trains with exponentially-decaying kernels and then sampling the resulting traces we obtain activity vectors with real-valued elements. A...
Digital Synesthesia, 2016
arXiv (Cornell University), Mar 26, 2023
Handbuch Data Science und KI
Neural activity in vitro can show bursts of activity. These bursts are termed neural avalanches a... more Neural activity in vitro can show bursts of activity. These bursts are termed neural avalanches and their size distribution f(s) approximates a power law [Beggs Plenz, 2003]. Since power law distributions are characteristic for self-organized critical (SOC) states [Bak, Tang, Wiesenfeld, 1987], neural activity was proposed to be SOC, too. Moreover, SOC may provide a basis for optimal information processing [Shew Plenz, 2013]. Evidence for the SOC hypothesis has been obtained for coarse measures of neural activity (LFP, EEG, MEG, BOLD), but surprisingly for spiking activity in vivo evidence for SOC is still missing. Therefore we analyzed highly parallel spike recordings from rats (hippocampus), cats (visual cortex) and monkeys (prefrontal cortex). For all recordings the f(s) were similar (Fig. 1 A-C), but showed fundamental differences to f(s) from a critical spiking model (Fig. 1 D), even under subsampling (Fig. 1 E). The differences between in vivo dynamics and model dynamics could be overcome by decreasing the model’s excitatory synaptic strength, while increasing its external input commensurately (Fig. 1 F). Thereby the model became subcritical and its separation of time scales (STS), which is fundamental to SOC, was eliminated. The match between the subcritical model and the neural activity held for standard and novel avalanche measures (f(s); branching parameter; mean avalanche size; frequency of single events) even when changing the temporal bin size over its full range. In addition, we showed that the same results held for local field potentials recorded in humans. These results suggest that neural activity in vivo is not SOC, but instead reflects a slightly subcritical regime without STS. This regime strikes a balance between optimal information processing and the need to avoid runaway activity. In this regime, avalanches are not temporally separated bursts, but form a mélange. Potential advantages of this regime compared to SOC are faster information processing (due to the lack of STS) and keeping a safety margin from supercriticality, which has been linked to epilepsy
Frontiers in Systems Neuroscience, 2019
Machine learning has achieved remarkable success with deep learning technologies. However, these ... more Machine learning has achieved remarkable success with deep learning technologies. However, these methods are often inefficient in terms of resources; they require large datasets, many parameters and consume much computational power. In this paper, I define a general strategy for machine learning, named strong machine learning, which aims to create resource-effective machine learning models. Under strong machine learning fall all the approaches that learn inductive biases during an initial phase and later apply those inductive biases to make models more effective learners. Several strong machine learning methods already exist and some are very popular exactly due to their effectiveness. However, strong machine learning is in its infancy and a lot more can be done. In order to further advance AI, we need to direct our effort toward developing even better, more powerful strong machine learning methods.
Computational Biology and Chemistry
The Handbook of Data Science and AI
Is clustering necessary? The mean rate classifier is applied directly on the spike-trains and the... more Is clustering necessary? The mean rate classifier is applied directly on the spike-trains and therefore is not affected by clustering. The trajectory classifier can also operate directly on the unclustered activity vectors because it relies just on Euclidean distances. However, the2 specificity classifier is dependent on clustering because it needs a finite set of patterns on which to compute statistics such as specificity and frequency of occurrence. Activity vectors cannot be directly used in this latter case and clustering must be performed to identify model vectors, because of the following reasons. Our approach (and others as well – see Baker and Gerstein [1]) does not binarize and bin multineuronal spike data, as is the case in other methods [2, 3] where a finite set of 2n vectors (n is the number of neurons) is obtained. By convolving spike trains with exponentially-decaying kernels and then sampling the resulting traces we obtain activity vectors with real-valued elements. A...
Digital Synesthesia, 2016
We are normally being told that it is the network of neurons that performs computations and that ... more We are normally being told that it is the network of neurons that performs computations and that these computations are the equivalent of cognitive operations (perception, attention, decision, memory recall, etc.). I have recently developed a new theory of intelligent adaptive systems, named "practopoiesis", and after applying the theory to the problem of brain-mind I came to a very surprising conclusion: The network computations cannot possibly account for cognition. Instead, the theory implies that it should be the local mechanisms of neural adaptation that are primarily responsible for generating mental operations. Thus, a thought is not a computation of a network. Rather, a thought is a process of adaptation of that network. In my talk I will present theoretical arguments and empirical evidence supporting these counterintuitive claims. I will also mention testable empirical predictions.
The traditional model of our mental function is that first our senses provide data to our brain, ... more The traditional model of our mental function is that first our senses provide data to our brain, which then translates those senses into the appropriate mental phenomena: light into visual images, air vibrations into auditory experiences, etc. But what if that process is actually occurring simultaneously? Danko Nikolić describes the theory of ideasthesia.
Afterimages result from a prolonged exposure to stillvisual stimuli. Theyare best detectable when... more Afterimages result from a prolonged exposure to stillvisual stimuli. Theyare best detectable when viewed against uniform backgroundsand can persist for multiple seconds. Consequently, the dynamics of afterimages appears to be slow bytheir verynature. To the contrary, we report here that about 50% of an afterimageintensitycan be erased rapidly—within less than a second. The prerequisite is that subjectsview a rich visual contentto erase the afterimage; fast erasure of afterimages does not occur if subjects view a blank screen. Moreover, we find evidence that fast removal of afterimages is a skilllearned with practiceasour subjects were always more effective in cleaning up afterimages in later parts of the experiment. These results can be explained by a tri-level hierarchy of adaptive mechanisms, as has been proposed by the theory of practopoiesis.
Ideasthesia can be defined as a phenomenon in which activation of concepts produces phenomenal ex... more Ideasthesia can be defined as a phenomenon in which activation of concepts produces phenomenal experience. The present article is concerned with the relationship between ideasthesia and art. In the past, it has proven difficult to come up with a comprehensive definition of art. Equally difficult seems to be to understand which psychological processes specifically underlie the creation and consumption of art. Here, an attempt is made to explain the psychology of art, as well as define art, based on the theory of ideasthesia. According to the present theory, art happens when the intensities of the meaning produced by a certain creation and the intensities of the experiences induced by that creation, are balanced out.