Sepp Hochreiter

is this you? claim profile

Johannes Kepler University Linz

Sepp Hochreiter is an informatician from Germany. Since 2018, after leading the Bioinformatics Institute from 2006 to 2018, he heads the Institute for Machine Learning at the Johannes Kepler University in Linz. He also heads the Linz Institute of Technology AI Lab since 2017, which is focused on promoting artificial intelligence research. He previously worked at Berlin University of Technology, Boulder University in Colorado and the University of Technology in Munich.

Sepp Hochreiter has contributed in numerous ways to master learning, deep learning and bioinformatics. He developed the long-term memory for which the first findings of his diploma thesis were reported in 1991. The main LSTM paper was published in 1997 as a discovery that is a milestone in machine learning. His analysis of the gradient disappearing or exploding led to the foundation of deep learning. He contributed to meta-learning and proposed flat minimum learning solutions to ensure the low generalization error. Artificial neural networks To improve learning, he has developed new activation functions for neural networks such as exponential linear units or scaled ELUs. He helped to strengthen learning through actor-critical approaches and his RUDDER method. He used drug discovery and toxicology biclustering techniques. He extended support for vector systems to handle kernels with a “Potential Support Vector Machine” model and used the model to select features, especially for genes for microarray data. In addition to his research papers, Sepp Hochreiter is widely involved in his field: he has started the Bioinformatics Working Group of the Austrian Society for Computers; he is a founding board member of several start-up bioinformatics companies; he was program chairman of the Bioinformatics conference Resea He founded the bioinformatics bachelor programme, a cross-border double-degree program, in conjunction with the University of South Bohemia in České Budějovice, Czech Republic, as a faculty member at Johannes Kepler Linz. He also established the Bioinformatics Masters Program,

Where the acting dean of both studies is still.

The long-term memory for which the first results were reported was developed in 1991 by Sepp Hochreiter in his diploma thesis. The main LSTM paper was published in 1997 and is considered a finding which is a landmark in machine learning. LSTM overcomes recurring neural networks and deep networks in order to forget information over time or equally over layers.

LSTM learns from training sequences in order to process new sequences to generate an output or an output sequence. LSTM cell neural networks solved numerous tasks in biological sequency analysis, drug design, automatic composition of music, machine translation, language recognition, improved learning and robotics.

LSTM with an optimized architecture has been used very quickly

Detection of protein homology without sequence alignment.

LSTM has been used to learn an algorithm, i.e. LSTM works as a Turing machine, i.e. a computer that performs an algorithm of learning. As the LSTM Turing machine is a neural network, new learning algorithms can be developed by learning about problems of learning. The new learning techniques that have been learnt are superior to those designed by humans. LSTM Networks are used for Google Voice, Google Voice Search and Google’s Allo as core technology in Google App’s voice searches and commands and for Android dictation. Since iOS 10, Apple has also used LSTM in its Quicktype function.

Neural networks are various types

Simplified biological neural network mathematical models like

Information only moves in one direction in feedforward neural networks,

from the input layer receiving environmental information,

The hidden layers to the output layer that provides environmental information.

In contrast to NNs, recurring neural networks

You can use your internal memory to process arbitrary input sequences.

When data mining is based on neural networks, overfitting reduces the network’s ability to process future data correctly. Sepp Hochreiter to avoid overfitting

Developed algorithms to locate neural networks of low complexity such as “Flat Minimum Search”

What looks for a “flat” minimum — a large connected area in the parameter space

The function of the network is constant. Network parameters can therefore be given with low accuracy

Means a low-complex, overfitting network. Low complexity neural network is well suited for deep learning, because it controls the complexity of each network layer and therefore learns to represent the input in a hierarchical way.

The group from Sepp Hochreiter introduced “explosive linear units” that accelerate learning in deep neural networks and lead to more accurate classification. Like corrected linear units, leaky ReLUs and parametric ReLUs, ELUs alleviate the problem of vanishing gradients by identifying positive values. However, compared to the ReLUs, ELUs have improved learning properties due to negative values that push unit activations closer to zero. Mean shifts to zero speed up learning by closer to the natural unit gradient due to a reduced bias-shift effect. Sepp Hochreiter has introduced neural self-normalization

Networks that provide feedforward networks with abstract input representations

SNNs avoid batch normalization problems since sample activation

Converge automatically to mean zero and variance one.

SNNs a technology that enables

train very deep networks with i.e.

use new strategies of regularization and

Learn in many layers very robustly. Unattended profound learning,

Generative opponent networks are very popular because they are

Create more realistic new images than those obtained from other generative approaches.

Sepp Hochreiter suggested

a two-time update rule for learning Stochastic gradient descent GANs

Any differentiable function of loss.

Stochastic approximation methods have been used to demonstrate

The TTUR is converging to a stationary local Nash balance.

This is the first evidence of the overall convergence of GANs.

The introduction of Another contribution is

the more appropriate “Fréchet Inception Distance”

Quality measurement for GANs than the Inception Score previously used. He has developed corrected factor networks

To build very sparse, non-linear, high-dimensional input representations efficiently. RFN models identify rare and small input events, have low interference between the units, have small rehabilitation errors and explain the structure of data covariance. RFN learning is a widespread alternative minimisation algorithm derived from a post regularization method that enforces non-negative and standard retroactive means. RFN has been applied to bioinformatics and genetics very successfully.

Sepp Hochreiter has been working on actor-critical systems in the field of strengthening learning

Learn by “model back propagation.” This approach, however, has major drawbacks

Analysis of sensitivity like

Local minima, various instabilities during online learning,

Explosion and disappearance of the world model gradients,

No contribution or relevance is assigned to the reward

“RUDDER: Return decomposition for delayed rewards,” introduced Sepp Hochreiter

which is intended to learn optimal policies with highly late rewards for Markov Decision Processes.

For retarded rewards, the biases of action-value assessments have been shown

The time difference is only slowly corrected exponentially

Number of steps of delay.

In addition, he has demonstrated that the variance of an action value estimate is

The Monte Carlo methods learned increase other estimates,

The number of which with the number of delay steps can grow exponentially.

Both the exponentially slow TD bias correction and the RUDDER solution

Increase of many MC variances by return decomposition exponentially.

A new MDP built by RUDDER returns for each episode and policy the same as the original

MDP, but the benefits are redistributed

Episode. Episode. The redistribution leads to significantly reduced reward delays.

In the best case, the new MDP does not have any delayed benefits and TD is

Unbiased. Unbiased. The redeployed rewards are designed to track Q values to

Always keep the expected reward for the future at zero. An action, therefore

Increasing the anticipated return receives a positive reward and

Action reducing the expected return receives a negative compensation.

A safe exploration strategy, a lesson for RUDDER

Replay buffer and a redistribution method based on LSTM

through decomposition of returns and analysis of backward contributions. Both source and source code

Videos of demonstration are available.

The exploration can be enhanced through active exploration strategies

Maximize future episodes information gain, often

Curiosity associated.

Many chemical compounds fail in the late phases of the drug development pipeline in the pharmaceutical industry. These failures are due to inadequate biomolecular target effectiveness, unwanted interactions with other biomolecules, or unforeseen toxic effects. Sepp Hochreiter’s profound learning and biclustering methods identified novel on- and off-target effects in various drug design projects. In 2013, the Group of Sepp Hochreiter won the DREAM subcontest to forecast the average compound toxicity. This success with Deep Learning was continued in 2014 by winning the NIH, FDA, and NCATS “Tox21 Data Challenge.” Tox21 Data Challenge’s objective was to correctly predict environmental chemicals’ off-target and toxic effects on nutrients, household products and drugs. These impressive successes demonstrate that profound learning is superior to other virtual screening methods. In addition, the Hochreiter Group worked to identify synergistic effects of drug combinations.

Sepp Hochreiter has developed “Bicluster Acquisition Factor Analysis” for biclustering which simultaneously clusters rows and matrix columns. A transcriptomic data bicluster is a pair of genes and a set of samples for which the genes are similar in the samples and vice versa. For example, the effects of compounds on a subgroup of genes may be similar in drug design. FABIA is a multiplicative model that takes on realistic non-Gussian signal distributions with heavy tails and uses well-understood models, like a Bayesian variational approach. FABIA supplies every bicluster with the information content to separate fake biclusters from true biclusters. Sepp Hochreiter published a reference book on biclustering which presents the most important algorithms, typical biclustering applications, bicluster visualization and evaluation and software in R.

Vector support machines are supervised methods of learning for

Analysis of classification and regression by recognizing patterns and data regularities. Standard SVMs demand a definite positive

Kernel to build a squared data kernel matrix. Sepp Hochreiter proposed the “Potential Support Vector Machine,” which can be used for non-square kernel matrices and with kernels which do not have a definite positive value. A efficient sequential minimum optimization algorithm has been developed for the PSVM model selection. The PSVM minimizes a new objective which ensures theoretical limitations of the generalization error and selects features which are used for classification or regression automatically.

Sepp Hochreiter used PSVM for selection of functions, in particular gene selection for microarray data.

The PSVM and standard vector support machines have been used to extract indicative features

Coiled oligomerization of the coil.

“HapFABIA: Identification of very short identity segments with descent characterized by rare variants in large sequence data,” was developed by Sepp Hochreiter in order to detect short identity segments with descent. A segment of DNA in two or more individuals is identical by state if the nucleotide sequences are identical in this segment. An IBS segment is identical in two or more descents

If they have inherited it from a common ancestor, the segment of these individuals has the same ancestral origin. HapFABIA identifies 100 times the smallest of IBD segments of state-of-the-art methods. For state-of-the-art methods 10kbp for HapFABIA vs 1Mbp. HapFABIA is adapted to next-generation sequencing data and uses rare variants for IBD detection as well as genotyping microarray data. HapFABIA enables evolutionary biology to be improved,

Population genetics and association trials because the genome has been decomposed in short, high-resolution IBD segments that describe the genome. HapFABIA was used to analyze human, neanderthal, and denisovan IBD sharing.

Sepp Hochreiter’s research group is part of the US Food and Drug Administration’s SEQC/MAQC-III Consortium. This consortium examined the RNA sequencing performance of Illumina HiSeq, Life Technologies SOLiD and Roche 454 platforms in multiple laboratory sites. This project defined standard approaches for evaluating, reporting and comparing the technical performance of experiments on differential gene expression at the genome scale. Sepp Hochreiter’s research group proposed to analyze the structural variation of DNA “cn.MOPS: Poissons mixture to identify copy number variations for next generation data at a low false discovery rate”

For detecting variations in copy number of next generation data sequence. The local DNA copy number is estimated by cn.MOPS, suitable for both whole genome sequencing and exom sequencing,

And may be applied to genomes diploid, haploid, and polyploid. Sepp Hochreiter’s group suggested “DEXUS: Identification of the Differential Expression in RNA-Seq Studies with Unknown Conditions” for the identification of differential transcripts in RNA-seq data. DEXUS can detect differential expression in RNA-seq data for which the sample is used.

Conditions that do not include biological replicates are unknown.

Sequence data were analyzed in the Sepp Hochreiter group to gain insights into chromatin remodeling. The The The

The cell structure reorganization was determined through the following generation sequencing of the resting and activated T cells. The analysis of this data on cell chromatin sequence identified long-lasting GC-rich

Nucleosome-free regions that are chromatin reshaping hot spots. For targeted clinical diagnostic next-generation sequence panels, particularly for cancer,

Panelcn.MOPS was developed by Hochreiter’s group.

Sepp Hochreiter has developed “Robust Microarray Summary Factor Analysis.” FARMS is designed to pre-treat and synthesize high-density DNA microarrays at the sample level for the analysis of the expression of RNA genes. FARMS is based on a factor analysis model optimized by maximizing the retro probability in a Bayesian context. FARMS outperformed all other methods with Affymetrix spiked in and other benchmark data. FARMS’ informative/non-informative calls are a highly relevant feature. The I/NI call is a Bayesian filtering method that divides signal variance from noise variance. In analyzing microarray data, the I/NI call offers a solution to the main problem of high dimensionality by selecting high quality measured genes. FARMS was expanded to cn.FARMS

Structural variants like copy number variations for the detection of DNA

With a low rate of false discovery.

Featured Co-authors