Lifeboat Foundation

How does a gambler maximize winnings from a row of slot machines? This question inspired the “multi-armed bandit problem,” a common task in reinforcement learning in which “agents” make choices to earn rewards. Recently, an international team of researchers, led by Hiroaki Shinkawa from the University of Tokyo, introduced an advanced photonic reinforcement learning method that transitions from the static bandit problem to a more intricate dynamic setting. Their findings were recently published in the journal, Intelligent Computing.

The success of the scheme relies on both a photonic system to enhance the learning quality and a supporting algorithm. Looking at a “potential photonic implementation,” the authors developed a modified bandit Q-learning algorithm and validated its effectiveness through numerical simulations. They also tested their algorithm with a parallel architecture, where multiple agents operate at the same time, and found that the key to accelerating the parallel learning process is to avoid conflicting decisions by taking advantage of the quantum interference of photons.

Although using the quantum interference of photons is not new in this field, the authors believe this study is “the first to connect the notion of photonic cooperative decision-making with Q-learning and apply it to a dynamic environment.” Reinforcement learning problems are generally set in a dynamic environment that changes with the agents’ actions and are thus more complex than the static environment in a bandit problem.

What happens when – instead of recording a single particle track or energy deposit in your detector – you see a complex collection of many particles, with many tracks, that leaves a large amount of energy in your calorimeters? Then congratulations: you’ve recorded a “jet”! Jets are the complicated experimental signatures left behind by showers of strongly-interacting quarks and gluons. By studying the internal energy flow of a jet – also known as the “jet substructure” – physicists can learn about the kind of particle that created it. For instance, several hypothesised new particles could decay into heavy Standard Model particles at extremely high (or “boosted”) energies. These particles could then decay into multiple quarks, leaving behind “boosted”, multi-pronged jets in the ATLAS experiment. Physicists use “taggers” to distinguish these jets from background jets created by single quarks and gluons. The type of quarks produced in the jet can also give extra information about the original particle. For example, Higgs bosons and top quarks often decay to b-quarks – seen in ATLAS as “b-jets” – which can be distinguished from other kinds of jets using the long lifetime of the B-hadron. The complexity of jets naturally lends itself to Artificial Intelligence (AI) algorithms, which are able to efficiently distil large amounts of information into accurate decisions. AI algorithms have been a regular part of ATLAS data analysis for several years, with ATLAS physicists continuously pushing these tools to new limits. This week, ATLAS physicists presented four exciting new results about jet tagging using AI algorithms at the BOOST 2023 conference held at Lawrence Berkeley National Lab (USA). Figure 1: The graphs showing the full declustering shower development and the primary Lund jet plane in red are shown in (left) for a jet originating from a W-boson and in (right) for a jet originating from a light-quark. (Image: ATLAS Collaboration/CERN) Artificial intelligence is revolutionising how ATLAS researchers identify – or ‘tag’ – what types of particles create jets in the experiment. Two results showcased new ATLAS taggers used for identifying jets coming from a boosted W-boson decay as opposed to background jets originating from light quarks and gluons. Typically, AI algorithms are trained on “high-level” jet substructure information recorded by the ATLAS inner detector and calorimeters – such as the jet mass, energy correlation ratios and jet splitting scales. These new studies instead use “low-level” information from these same detectors – such as the direct kinematic properties of a jet’s constituents or the novel two-dimensional parameterisation of radiation within a jet (known as the “Lund Jet plane”), built from the jet’s constituents and using graphs based on the particle-shower development (see Figure 1). These new taggers made it possible to separate the shape of signal and background far more effectively than any high-level taggers could do alone (see Figure 2). In particular, the Lund Jet plane-based tagger outperforms the other methods, by using the same input to the AI networks but in a different format inspired by the physics of the jet shower development. A similar evolution was followed for the development of a new boosted Higgs tagger, which identifies jets originating from boosted Higgs bosons decaying hadronically to two b-quarks or c-quarks. It also uses low-level information – in this case, tracks reconstructed from the inner detector associated with the single jet containing the Higgs boson decays. This new tagger is the most performant tagger to date, and represents a factor of 1.6 to 2.5 improvement, at a 50% boosted Higgs signal efficiency, over the previous version of the tagger, which used high-level information from the jet and b/c-quark decays as input for a neural network (see Figure 3). Figure 2: Signal efficiency as a function of the background rejection for the different W-boson taggers: one is based on the Lund jet plane, while the others use unordered sets of particles or graphs with additional structure. (Image: ATLAS Collaboration/CERN) Figure 3: Top and multijet rejections as a function of the H→bb signal efficiency. Performance of the new boosted Higgs tagger is compared to the previous taggers using high-level information from the jet b-quark decays. (Image: ATLAS Collaboration/CERN) Finally, ATLAS researchers presented two new taggers that aim to differentiate between jets originating from quarks and those originating from gluons. One tagger looked at the charged-particle constituent multiplicity of the jets being tagged, while the other combined several jet kinematic and jet substructure variables using a Boosted Decision Tree. Physicists compared the performance of these quark/gluon taggers; Figure 4 shows the rejection of gluon jets as a function of quark selection efficiency in simulation. Several studies of Standard-Model processes – including vector boson fusion – and new physics searches with quark-rich signals could greatly benefit from these taggers. However, in order for them to be used in analyses, additional corrections on the signal efficiency and background rejection need to be applied to bring the performance of the taggers in data and simulation to be the same. Researchers measured both the efficiency and rejection rates in Run-2 data for these taggers, and found good agreement between the measured data and predictions; therefore, only small corrections are needed. The excellent performance of these new jet taggers does not come without questions. Crucially, how can researchers interpret what the machine-learning models learned? And why do more complex architectures show a stronger dependence on the modelling of simulated physics processes used for the training, as shown in the two W-tagging studies? Challenges aside, these taggers set an outstanding baseline for analysing LHC Run-3 data. Given the current strides being made in machine learning, its continued application to particle physics will hopefully increase the understanding of jets and revolutionise the ATLAS physics programme in the years to come. Figure 4: Signal efficiency as a function of the background rejection for different quark taggers. The use of machine learning (BDT) results in an improved performance. (Image: ATLAS Collaboration/CERN) Learn more Tagging boosted W bosons with the Lund jet plane in ATLAS (ATL-PHYS-PUB-2023–017) Constituent-based W-boson tagging with the ATLAS detector (ATL-PHYS-PUB-2023–020) Transformer Neural Networks for Identifying Boosted Higgs Bosons decaying into bb and cc in ATLAS (ATL-PHYS-PUB-2023–021) Performance and calibration of quark/gluon-jet taggers using 140 fb−1 of proton–proton collisions at 13 TeV with the ATLAS detector (JETM-2020–02) Comparison of ML algorithms for boosted W boson tagging (JETM-2023–003) Summary of new ATLAS results from BOOST 2023, ATLAS News, 31 July 2023.

A tailsitter is a special kind of fixed-wing aircraft that sits on its tail when it is on the ground and then tilts horizontally for forward flight.

In the ever-evolving world of aerial technology, MIT’s researchers have given wings to the brilliance of aircraft design with their new algorithms for tailsitter drones. This breathtaking technology is enabling these aircraft to execute astounding acrobatics and challenging maneuvers, paving the way for futuristic applications in search-and-rescue, parcel delivery, and more.

Continue reading “How MIT researchers made tailsitter drones fly like acrobats” »

Enter AI. Multiple deep learning methods can already accurately predict protein structures— a breakthrough half a century in the making. Subsequent studies using increasingly powerful algorithms have hallucinated protein structures untethered by the forces of evolution.

Yet these AI-generated structures have a downfall: although highly intricate, most are completely static—essentially, a sort of digital protein sculpture frozen in time.

A new study in Science this month broke the mold by adding flexibility to designer proteins. The new structures aren’t contortionists without limits. However, the designer proteins can stabilize into two different forms—think a hinge in either an open or closed configuration—depending on an external biological “lock.” Each state is analogous to a computer’s “0” or “1,” which subsequently controls the cell’s output.

Researchers from The University of Queensland applied an algorithm from a video game to study the dynamics of molecules in living brain cells.

Dr. Tristan Wallis and Professor Frederic Meunier from UQ’s Queensland Brain Institute came up with the idea while in lockdown during the COVID-19.

First identified in 2019 in Wuhan, China, COVID-19, or Coronavirus disease 2019, (which was originally called “2019 novel coronavirus” or 2019-nCoV) is an infectious disease caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). It has spread globally, resulting in the 2019–22 coronavirus pandemic.

Brain age was estimated using an algorithm that combined multiple measures of brain structure obtained through MRI scans when the participants were 45 years old. This algorithm quantified the difference between estimated brain age and the participants’ chronological age, referred to as brain age gap estimate.

If the estimated brain age is higher than the chronological age, it suggests that the brain’s structural characteristics are more similar to those of an older individual. Conversely, if the estimated brain age is lower than the chronological age, the brain’s structural characteristics resemble those of a younger individual.

Lay-Yee and his colleagues also adjusted their analyses for various potential confounding factors. These included socio-demographic factors like sex and socio-economic status, as well as family factors (teen-aged mother, single parent, change in residence, maltreatment) and child-behavioral factors (self-control, worry/fearfulness).

The development of robotic avatars could benefit from an improvement in how computers detect objects in low-resolution images.

A team at RIKEN has improved computer vision recognition capabilities by training algorithms to better identify objects in low-resolution images. Inspired by human brain memory formation techniques, the model degrades the quality of high-resolution images to train the algorithm in self-supervised learning, enhancing object recognition in low-quality images. The development is expected to benefit not only traditional computer vision applications but also the creation of cybernetic avatars and terahertz imaging technology.

Robotic avatar vision enhancement inspired by human perception.

A tiny little winged drone can soar with close to zero throttle, thanks to an impressive algorithm that responds to changing winds.

Ticking clocks and flashing fireflies that start out of sync will fall into sync, a tendency that has been observed for centuries. A discovery two decades ago therefore came as a surprise: the dynamics of identical coupled oscillators can also be asynchronous. The ability to fall in and out of sync, a behavior dubbed a chimera state, is generic to identical coupled oscillators and requires only that the coupling is nonlocal. Now Yasuhiro Yamada and Kensuke Inaba of NTT Basic Research Laboratories in Japan show that this behavior can be analyzed using a lattice model (the XY model) developed to understand antiferromagnetism [1]. Besides a pleasing correspondence, Yamada and Inaba say that their finding offers a path to study the partial synchronization of neurons that underlie brain function and dysfunction.

The chimera states of a system are typically analyzed by looking at how the relative phases of the coupled oscillators fall in and out of sync. But that approach struggles to describe the system when the system contains distantly separated pockets of synchrony or when there are nontrivial configurations of the oscillators, such as twisted or spiral waves. It also requires knowledge of the network’s structure and the oscillators’ equations of motion.

In seeking an alternative approach, Yamada and Inaba turned to a two-dimensional lattice model used to tackle phase transitions in 2D condensed-matter systems. A crucial ingredient in that model is a topological defect called a vortex. Yamada and Inaba found that they could embody the asynchronous dynamics of pairs of oscillators by formulating the problem in terms of an analogous quantity that they call pseudovorticity, whose absence indicates synchrony and whose presence indicates asynchrony. Their calculations show that their pseudo-vorticity-containing lattice model can successfully recover the chimera state behavior of a simulated neural network made up of 200 model oscillators of a type commonly used to study brain activity.